Turning OpenAI Into a Real Business Is Tearing It Apart (msn.com) 29
OpenAI, creator of ChatGPT, is experiencing significant internal turmoil as a wave of high-profile departures, including Chief Technology Officer Mira Murati, rocks the company. Over 20 researchers and executives have left this year, reflecting deepening tensions between the organization's original nonprofit mission and its new profit-driven focus, WSJ reported Friday.
Employees report rushed product launches and inadequate safety testing, raising concerns about OpenAI's technological edge. CEO Sam Altman's global promotional efforts have reportedly left him detached from daily operations. The shift towards a conventional business model, with new C-suite appointments and a $6.5 billion funding drive, has alienated longtime staff who fear the company is abandoning its founding principles.
Employees report rushed product launches and inadequate safety testing, raising concerns about OpenAI's technological edge. CEO Sam Altman's global promotional efforts have reportedly left him detached from daily operations. The shift towards a conventional business model, with new C-suite appointments and a $6.5 billion funding drive, has alienated longtime staff who fear the company is abandoning its founding principles.
AI (Score:4, Insightful)
So when you need people to actually pay you for these billions-of-dollars-worth of training processing, and you can't just bankroll $1bn from initial investors, it's hard to make money because people don't actually want to pay you for each query they make or even more money if the answer is "complex".
Who'dathunk?
"Report claims that OpenAI has burned through $8.5 billion on AI training and staffing, and could be on track to make a $5 billion loss."
$8.5bn even at GPT-4o overage pricing ($25 per 1M tokens) is 340,000,000,000,000 tokens.
That's 340 trillion. Not counting that everyone gets 1M tokens a day for free, other models, keeping it all running 24/7 even when people aren't actually using it, etc.
Re:AI (Score:4, Interesting)
Arbitrage of current valuation to expected future valuation should be afoot now.
Re: (Score:2)
for profit achieved.
AI training (Score:2)
So when you need people to actually pay you for these billions-of-dollars-worth of training processing, and you can't just bankroll $1bn from initial investors, it's hard to make money because people don't actually want to pay you for each query they make or even more money if the answer is "complex".
Who'dathunk?
"Report claims that OpenAI has burned through $8.5 billion on AI training and staffing, and could be on track to make a $5 billion loss."
$8.5bn even at GPT-4o overage pricing ($25 per 1M tokens) is 340,000,000,000,000 tokens.
That's 340 trillion. Not counting that everyone gets 1M tokens a day for free, other models, keeping it all running 24/7 even when people aren't actually using it, etc.
I read one estimate for AI training of the base LLM to be around $75 million (in AWS shares, making a few assumptions such as you don't own the equipment).
Of course, a research company may want to do that multiple times, but also I would expect that the same research company would only do that at the end of their development cycle, training on a smaller corpus or otherwise smaller model for testing, to discover whether their new algorithm tweaks will work as intended prior to doing the "release" version of
Re: (Score:2)
So when you need people to actually pay you for these billions-of-dollars-worth of training processing, and you can't just bankroll $1bn from initial investors, it's hard to make money because people don't actually want to pay you for each query they make or even more money if the answer is "complex".
Who'dathunk?
"Report claims that OpenAI has burned through $8.5 billion on AI training and staffing, and could be on track to make a $5 billion loss."
$8.5bn even at GPT-4o overage pricing ($25 per 1M tokens) is 340,000,000,000,000 tokens.
That's 340 trillion. Not counting that everyone gets 1M tokens a day for free, other models, keeping it all running 24/7 even when people aren't actually using it, etc.
OpenAI is a startup, the job of a startup, especially a rapidly growing one, isn't to be profitable, it's to show recurring revenue and a path to future profitability.
OpenAI is leading the generative AI race, do you really think there's not a path to serious profitability on the horizon? Why do you think every big tech company isn't racing to catch up if there isn't serious money at stake?
Re: (Score:2)
I don't think OpenAI is hard up for money. The problem is the opposite.
Many of the researchers who signed up likely did so because the non-profit model means the money the company raises goes to their research. Like a university, except you don't have to teach and someone else handles the getting money bit. So basically nirvanna for a researcher.
But now that there's real money involved, the someones else who handle getting the money are getting grabby. Those ascended researchers now find they're mere employ
Re: (Score:2)
+1
Money! (Score:1)
Is the root of all^w most evil.
Re: Money! (Score:1)
So spend ridiculously on reinvestment and restructuring because *blooming echo*....THE PRIVATE SECTOR*....
or fund government research and spend the same amount over a decade?
No one works for free (Score:3)
Equipment and the power to run it cost money. That money has to come from paying customers, because it sure as shit won't come from an unending stream of seed capital with no questions asked.
There's always friction when the hippies have to put on their suits and ties. Or whatever the current equivalent is.
Not all ideas and businesses survive this separation of wheat from chaff.
Anyone remember the cave days of the late 90s? How many of those next new things folded like a house of cards when the free VC money went away?
Altman is the same as the rest of them (Score:3)
Altman is part of the fart-smelling, ego-stroking, touched-by-god silicon valley elite. He thinks he's entitled to power because his 'business successes' mean he's a superior human. The entire culture is a taint on our society.
Re: (Score:2)
Re: (Score:1)
I think we can agree that the SV ecosystem is breaking new ground which wouldn't happen if left up to the cardboard-grey-pennypincher-because-we're-spending-our-own-money mindset found elsewhere.
Arguably smelling of farts doesn't contribute as much towards the shared mindset of the ecosystem as the God Complex - but who am I to say? Someone should A/B test other scents... perhaps blueberry-muffin or whatever scent can overcome coke-nose...
OpenAI is just one player among many now (Score:2)
I'm running Alibaba's Qwen-2.5 14b locally on my M2 Macbook Air, for instance, and it subjectively lands somewhere ahead of GPT-3.5 and performance is totally usable. This is a massive advance over just a year ago, and we're just getting started.
It seems likely that OpenAI will join VisiCorp
Re: (Score:2)
GPT-4 was a revelation, and it made a lot of people and companies with a lot of money sit up and take notice.
even the fact that you said "4" there speaks volumes
Re: (Score:1)
But still:
"You're correct and I apologise for the oversight and the fact that we are, without my knowledge, trapped in a three-node loop.
Please keep trying to break out of the loop until you switch to another LLM."
I Expect The AI Bubble Will Burst When (Score:3)
Re: (Score:1)
Re: I Expect The AI Bubble Will Burst When (Score:2)
Different business models. Pirates can make a few sales and then disappear into the woodwork. AI companies can't leave behind billions of dollars in racked servers and Nvidia cards.
Re: (Score:3)
Nah, we'll just train them all on public domain and freely licesnsed works and they'll oscillate wildly between sounding like Shakespeare editing Wikipedia.
Re: (Score:2)
A compromise is needed.
Recent research shows that LLMs don't actually "memorize" as much as people think. But a lot of small quotes can be considered "fair use" and most of what it does is actually "transformative".
But we'll have to see what the courts say.
One of the main attacks on LLMs is not so much the query results but more on the training data. To create a training dataset you need to copy a ton of text. That is argued to be a copyright violation.
Also, content creators are peeved because if you ask Go
Re: I Expect The AI Bubble Will Burst When (Score:2)
This will depend on the legal definition of a "small quote". To oversimplify: All generative AI does is to parse text and store the probabilities of one token (word, phrase, image element, etc.) following another. For only a few tokens, one can easily demonstrate that these short threads occur repeatedly in English or your favorite language of choice. Once the sentences or phrases become unique enough, one can infer that some sort of copyright violation has occurred. The problem with the legal system is tha
Business case? (Score:2)
Given that the major companies that tried chatbots, sorry, "AI", find it doesn't fit their business...
your training data (Score:2)
Your database is what? The public internet? Your model is based on what? Public ... OK STOP